Kyle Daruwalla
Kyle Daruwalla
I found a potential use case for this: anything stored in the phase struct can't be scheduled as a hyper-parameter. Either hyper-parameters should be extended to include the phase or...
I accessed the value history from one of my training runs. Here is plot. It looks clean, so I think the issue is limited to just logging.
It looks like `resnet[1:end-2]` is producing an empty `Chain` in your model. That's why taking the gradient throws an error. Can you post the cells that generate `resnet`?
Ah looks like the tutorial is outdated. Metalhead splits the model into a 2-element Chain of sub-Chains. The first element is the backbone (that you want for transfer learning), and...
Maybe a white outline on the arrow is better than a totally white arrow?
Yeah I think implementing the distance part of it in Distances.jl makes a lot of sense.
Yeah I think the hierarchy would be Distances -> Metrics -> Losses. There will be losses that are not metrics (i.e. defined completely in a loss package), and losses that...
@lorenzoh do you want to transfer stuff from here to FastAI.jl issues?
> P.S. there really needs to be a way to just run docs previews if it's only a docs change - the Metalhead CI load is quite heavy Separate it...
Is it affecting queuing? As long as we can get machines...it's GH's load not ours 😄.